Aug 28 (Reuters) – Smartphones and social media have transformed journalism by producing a torrent of photos and videos of almost anything that happens where people are present, from holidays to hurricanes, celebrations to crises.
The swift rise of AI has tainted that flow, mixing in highly realistic video and still images of events that never happened or, more perniciously, alternate versions of things that did happen but not in the way depicted.
Reuters confronts this reality with a team of journalists specializing in “visual verification,” who search the internet for images of breaking news events – from battlefields to border crossings – and then go through the painstaking process of confirming that those images do depict real events in the times and places they occurred.
Here’s a look at how Reuters journalists tackle this challenge head-on, sorting through hundreds of photos and videos per day, of which it might validate and publish a dozen:
WHY DOES REUTERS RELY ON IMAGES FROM THE PUBLIC?
As the world’s largest news agency, Reuters has an enormous footprint, with about 2,600 journalists spread over some 200 locations. But the Earth is a vast place and Reuters journalists can’t be everywhere – news happens every day at unexpected times and in unexpected places. Videos and photos taken by people who are eyewitnesses to history, verified by Reuters journalists, have become an indispensable part of the file produced by Reuters.
This is in line with the Reuters Trust Principles, which have guided the news agency since World War Two and include directives to provide unbiased, reliable news and to continually develop and strengthen its services.
Verified images sourced from the public have been critical in breaking major news stories from around the world, including providing evidence suggesting that a U.S. missile had been used in a strike on an Iranian girls’ school that killed scores of children and in providing detailed looks at the fatal Minneapolis shootings of American citizens protesting President Donald Trump’s immigration crackdown.
WHAT IS ARTIFICIAL INTELLIGENCE CAPABLE OF?
AI’s ability to generate realistic videos, photos and audio has grown increasingly sophisticated over the past few years. While the fabricated or manipulated images it produced were once littered with visual giveaways – people with unusual numbers of fingers, background text written in unrecognizable letters – what it produces now is far more life-like.
Beyond the technological advances, people who want to use AI to create misleading imagery can feed data into it – photos of real-world people, places and events that the AI can then alter to produce far more convincing images.
Following the U.S. capture of Venezuelan President Nicolás Maduro in January, some social media users circulated AI-generated images of Maduro in handcuffs – creating a false image of a real event. Reuters has also reported on AI-generated ads used to provide misleading images about candidates in the 2026 U.S. midterm elections.
Even without AI, some social media users will share photos or videos of past events but inaccurately label what they represent – such as taking videos of protests that did occur at a particular place and time in the past and describing them as happening in the present and at other locations – in response to other events.
SO HOW TO TELL FACT FROM FICTION?
The Reuters global visual verification team follows a systematic process to assess whether images found on social media are real, unaltered versions of the events that they purport to portray.
There are a few critical steps in the process. One is to find the person who took the photo or video, confirm that they are who they say they are and if possible interview them to learn more about their experience.
Many images also carry metadata that can confirm where and when they were made and on what device; if that is present it can play an important role in verification.
Reuters journalists on the verification team also compare what a photo or video shows to other known publicly available data and imagery. These include weather reports, satellite imagery, archive and street view imagery. The direction of shadows can reveal the time of day at which a photo or video was taken. Media and official reports on events are taken into account when trying to establish what happened. So too are corroborating images and videos taken by other eyewitnesses that may show the same scene from a different angle.
Visual verification journalists compare their work to assembling a puzzle — once the full picture can be seen, then it is time to publish. Each photo and video is also scanned with several AI detection tools that are trained to look for evidence of AI alteration or generation. These tools are not foolproof – at times our VVT team isn’t able to draw firm conclusions from the clues they yield – but such filters are important in spotting traces of AI that may not be visible to the naked eye.
The journalists inside the Reuters newsroom make a judgment call, which gets more challenging by the day.
(Writing by Scott Malone; Editing by Howard Goller)
Comments